+#endif /* !PIPE_ARCH_SSSE3 */
+
+
+/*
+ * Provide an SSE implementation of _mm_mul_epi32() in terms of
+ * _mm_mul_epu32().
+ *
+ * Basically, albeit surprising at first (and second, and third...) look
+ * if a * b is done signed instead of unsigned, can just
+ * subtract b from the high bits of the result if a is negative
+ * (and the same for a if b is negative). Modular arithmetic at its best!
+ *
+ * So for int32 a,b in crude pseudo-code ("*" here denoting a widening mul)
+ * fixupb = (signmask(b) & a) << 32ULL
+ * fixupa = (signmask(a) & b) << 32ULL
+ * a * b = (unsigned)a * (unsigned)b - fixupb - fixupa
+ * = (unsigned)a * (unsigned)b -(fixupb + fixupa)
+ *
+ * This does both lo (dwords 0/2) and hi parts (1/3) at the same time due
+ * to some optimization potential.
+ */
+static inline __m128i
+mm_mullohi_epi32(const __m128i a, const __m128i b, __m128i *res13)
+{
+ __m128i a13, b13, mul02, mul13;
+ __m128i anegmask, bnegmask, fixup, fixup02, fixup13;
+ a13 = _mm_shuffle_epi32(a, _MM_SHUFFLE(2,3,0,1));
+ b13 = _mm_shuffle_epi32(b, _MM_SHUFFLE(2,3,0,1));
+ anegmask = _mm_srai_epi32(a, 31);
+ bnegmask = _mm_srai_epi32(b, 31);
+ fixup = _mm_add_epi32(_mm_and_si128(anegmask, b),
+ _mm_and_si128(bnegmask, a));
+ mul02 = _mm_mul_epu32(a, b);
+ mul13 = _mm_mul_epu32(a13, b13);
+ fixup02 = _mm_slli_epi64(fixup, 32);
+ fixup13 = _mm_and_si128(fixup, _mm_set_epi32(-1,0,-1,0));
+ *res13 = _mm_sub_epi64(mul13, fixup13);
+ return _mm_sub_epi64(mul02, fixup02);
+}
+
+
+/* Provide an SSE2 implementation of _mm_mullo_epi32() in terms of
+ * _mm_mul_epu32().
+ *
+ * This always works regardless the signs of the operands, since
+ * the high bits (which would be different) aren't used.
+ *
+ * This seems close enough to the speed of SSE4 and the real
+ * _mm_mullo_epi32() intrinsic as to not justify adding an sse4
+ * dependency at this point.
+ */
+static inline __m128i mm_mullo_epi32(const __m128i a, const __m128i b)
+{
+ __m128i a4 = _mm_srli_epi64(a, 32); /* shift by one dword */
+ __m128i b4 = _mm_srli_epi64(b, 32); /* shift by one dword */
+ __m128i ba = _mm_mul_epu32(b, a); /* multply dwords 0, 2 */
+ __m128i b4a4 = _mm_mul_epu32(b4, a4); /* multiply dwords 1, 3 */
+
+ /* Interleave the results, either with shuffles or (slightly
+ * faster) direct bit operations:
+ * XXX: might be only true for some cpus (in particular 65nm
+ * Core 2). On most cpus (including that Core 2, but not Nehalem...)
+ * using _mm_shuffle_ps/_mm_shuffle_epi32 might also be faster
+ * than using the 3 instructions below. But logic should be fine
+ * as well, we can't have optimal solution for all cpus (if anything,
+ * should just use _mm_mullo_epi32() if sse41 is available...).
+ */
+#if 0
+ __m128i ba8 = _mm_shuffle_epi32(ba, 8);
+ __m128i b4a48 = _mm_shuffle_epi32(b4a4, 8);
+ __m128i result = _mm_unpacklo_epi32(ba8, b4a48);
+#else
+ __m128i mask = _mm_setr_epi32(~0,0,~0,0);
+ __m128i ba_mask = _mm_and_si128(ba, mask);
+ __m128i b4a4_mask_shift = _mm_slli_epi64(b4a4, 32);
+ __m128i result = _mm_or_si128(ba_mask, b4a4_mask_shift);
+#endif
+
+ return result;
+}
+
+
+static inline void
+transpose4_epi32(const __m128i * restrict a,
+ const __m128i * restrict b,
+ const __m128i * restrict c,
+ const __m128i * restrict d,
+ __m128i * restrict o,
+ __m128i * restrict p,
+ __m128i * restrict q,
+ __m128i * restrict r)
+{
+ __m128i t0 = _mm_unpacklo_epi32(*a, *b);
+ __m128i t1 = _mm_unpacklo_epi32(*c, *d);
+ __m128i t2 = _mm_unpackhi_epi32(*a, *b);
+ __m128i t3 = _mm_unpackhi_epi32(*c, *d);
+
+ *o = _mm_unpacklo_epi64(t0, t1);
+ *p = _mm_unpackhi_epi64(t0, t1);
+ *q = _mm_unpacklo_epi64(t2, t3);
+ *r = _mm_unpackhi_epi64(t2, t3);
+}
+
+
+/*
+ * Same as above, except the first two values are already interleaved
+ * (i.e. contain 64bit values).
+ */
+static inline void
+transpose2_64_2_32(const __m128i * restrict a01,
+ const __m128i * restrict a23,
+ const __m128i * restrict c,
+ const __m128i * restrict d,
+ __m128i * restrict o,
+ __m128i * restrict p,
+ __m128i * restrict q,
+ __m128i * restrict r)
+{
+ __m128i t0 = *a01;
+ __m128i t1 = _mm_unpacklo_epi32(*c, *d);
+ __m128i t2 = *a23;
+ __m128i t3 = _mm_unpackhi_epi32(*c, *d);
+
+ *o = _mm_unpacklo_epi64(t0, t1);
+ *p = _mm_unpackhi_epi64(t0, t1);
+ *q = _mm_unpacklo_epi64(t2, t3);
+ *r = _mm_unpackhi_epi64(t2, t3);
+}
+
+
+#define SCALAR_EPI32(m, i) _mm_shuffle_epi32((m), _MM_SHUFFLE(i,i,i,i))
+